Search

A case study of CNN, a spatial deep learning model, is presented in this chapter for analyzing water quality with remote sensing data. Airborne remote sensing of cyanobacteria with multispectral/hyperspectral sensors is employed for input data and its water quality as output data.

A neural network is composed of a number of parameters, also called weights. The neural network training is the estimation procedure of those parameters. In this chapter, the training procedure of a neural network is described based on the gradient descent method and the backpropagation algorithm.

In order improve the performance of a neural network model, a number of ways have been studied. In this chapter, minibatch and k-fold cross-validation are explained. The basic idea of these two methods is on controlling the dataset, since repeated usage of the same dataset for training and validation might result in overfitting. Furthermore, regularization of the neural network model training by L-norm regularization and dropout of hidden nodes are explained in this chapter to avoid overfitting.

Tensorflow is an end-to-end open-source platform for machine learning containing a comprehensive, flexible ecosystem of tools, libraries, and community resources (https://www.tensorflow.org/). It provides multiple levels of abstractions to choose the right one. The high-level Keras API can be used to build and train models by easily getting started with Tensorflow. Keras allows employing Tensorflow without losing its flexibility and capability. In the following, two applications (i.e., temporal and spatial deep learning) are presented to illustrate how to use Keras with python.

One of the major applications in deep learning models is to forecast the future. In recent years, time series forecasting with deep learning models has been developed and applied in a number of fields. Recurrent neural network models can allow forecasting future better, and long short-term memory (LSTM) is a breakthrough to overcome the shortages of the previous RNN model. These algorithms are explained in detail in this chapter.

Among recent developments of deep learning models, the availability of spatial datasets or images with deep learning is the most significant contribution. In this chapter, a convolutional neural network (CNN) that can analyze the spatial datasets is described. The training procedure of CNN is described with a simple example.

Deep learning models have been applied for hydrometeorological applications. In this chapter, a few of them are explained. In the field of hydrometeorology, time-series deep learning models are mainly employed. In this chapter, the development procedure of a time series deep learning model for stochastic simulation producing a long sequence that mimics historical series is explained. Furthermore, the case study for daily maximum temperature with an LSTM model is presented.